Your agents need knowledge,
not another database

What is a knowledge graph?

Nodes are facts — concepts, operators, frameworks, cases. Edges are relationships — teaches, extends, illustrates. Confidence encodes trust. Click any node to see how traversal discovers connected knowledge.

Node color = type · brightness = confidence · click to traverse

Not a graph database. A knowledge engine.

Graph databases store and query relationships. h2t-graphs goes further — schema-guided extraction, enrichment pipeline, quality gates, and LLM-optimized output. Your agent gets answers, not rows.

Engine owns infrastructure. Domain owns knowledge.

3-layer ownership model. The engine knows nothing about Houdini, TouchDesigner, or creative thinking. It stores, scores, queries, and validates. Consumers bring domain data and get domain answers.

graph LR
    subgraph CLIENT ["CLIENT LAYER"]
        D1["h2t-ai
DCC Operators"] & D2["creative-thinking
Methodology"] & D3["h2t-transcription
Courses"]
    end
    subgraph ENGINE ["h2t-graphs ENGINE"]
        E1["Store"] & E2["Score"] & E3["Search"] & E4["Validate"] & E5["Enrich"]
    end
    subgraph STUDIO ["GRAPH STUDIO (future)"]
        S1["Templates"] & S2["LLM Copilot"] & S3["ETL Pipeline"]
    end
    CLIENT -->|"data + feedback"| ENGINE
    ENGINE -->|"answers"| CLIENT
    STUDIO -->|"schema + enrichment"| ENGINE
    style ENGINE fill:#1a0a10,stroke:#e94560,color:#e94560
    style CLIENT fill:#0a1a0d,stroke:#00ff88,color:#c0c0d0
    style STUDIO fill:#0a0d1a,stroke:#4a9eff,color:#c0c0d0

Currently serving: DCC operators (4,363), creative methodology (901), course transcripts (9,379), TRIZ principles, design thinking

Schema-guided enrichment. Six phases.

Template-first approach: strict typed constraints → pre-filter invalid pairs → guided reasoning LLM → judge with traceability → quality gate. Without template constraints, LLM accuracy = 40%. With template + guided reasoning = 68%, target 85%+.

Four search modes. One API call.

Keyword precision, semantic understanding, or both. Edge-enriched embeddings capture graph context, not just node text.

Every node has a trust score

Not all knowledge is equal. LLM-generated content starts at low confidence. Expert-verified content earns trust. The engine tracks the journey.

graph LR
    M["MODEL
confidence: 0.3"] -->|"human review"| S["SESSION
confidence: 0.6"]
    S -->|"expert validation"| E["EXPERT
confidence: 0.9"]
    E -->|"feedback refines"| S
    style M fill:#0e0e14,stroke:#3a3a50,color:#a0a0b8
    style S fill:#0a0d1a,stroke:#4a9eff,color:#4a9eff
    style E fill:#0a1a0d,stroke:#00ff88,color:#00ff88,stroke-width:2px

Source types: expert (0.8–1.0) · docs (0.9) · session (0.6–0.8) · model (0.3–0.5). Queries filter by min_confidence. Feedback adjusts scores in real time.

Write a node. Search it instantly.

On-write embedding — every new node is automatically vectorized with edge context. No batch jobs, no waiting. The node is semantically searchable the moment it's written.

# Write a node via API
POST /api/nodes
{
  "source": "my-domain",
  "node": {"id": "insight_42", "node_type": "insight", "label": "Pattern discovered", ...}
}

→ {"status": "ok", "node_id": "insight_42", "embedded": true}

# Immediately searchable
GET /api/query?source=my-domain&semantic=pattern+discovery
→ results include insight_42

Why schema-guided matters

Without schema constraints, LLM extraction produces 40–76% accuracy. Schema-guided extraction with typed constraints changes the game.

Differentiator: Schema-guided extraction as default (not optional), with provenance, enrichment pipeline, and quality gates in one package.

By the Numbers

Part of a self-improving ecosystem

h2t-graphs is the persistence layer in a larger feedback loop. Evaluation, optimization, and continuous learning flow through the graph.

graph LR
    subgraph CLIENTS ["CLIENTS"]
        C1["DCC Skill"] & C2["Creative"] & C3["Transcription"]
    end
    subgraph GRAPHS ["h2t-graphs"]
        G1["Store"] & G2["Search"] & G3["Enrich"] & G4["Score"]
    end
    subgraph EVALS ["h2t-evals"]
        V1["GT Packs"] & V2["Judge"] & V3["Optimizer"]
    end
    CLIENTS -->|"query + feedback"| GRAPHS
    GRAPHS -->|"metrics"| EVALS
    EVALS -->|"improve"| CLIENTS
    style GRAPHS fill:#1a0a10,stroke:#e94560,color:#e94560
    style CLIENTS fill:#0a1a0d,stroke:#00ff88,color:#c0c0d0
    style EVALS fill:#0a0d1a,stroke:#4a9eff,color:#c0c0d0

What's next

From template-first enrichment to a full Graph Studio with LLM copilot.

Built for solo + AI teams

Lightweight. No Kubernetes. No managed services. One VPS, auto-TLS, full control. 601 tests. CI/CD: push main → deploy.